Empirical evaluation of scoring functions for Bayesian network model selection.

Area of Science:

• Computational statistics
• Machine learning
• Data mining

Background:

• Previous studies on Bayesian network structure learning often used approximate algorithms, yielding suboptimal results.
• Prior research utilized synthetic networks, limiting applicability to real-world scenarios.

Purpose of the Study:

• To empirically evaluate Bayesian network scoring functions for accurate structure recovery.
• To address limitations of prior work by using optimal learning algorithms and real-world data.

Main Methods:

• Learned Bayesian network structures using an optimal algorithm on datasets derived from real-world gold-standard networks.
• Compared performance of Minimum Description Length (MDL/BIC), Akaike's Information Criterion (AIC), Bayesian Dirichlet equivalence score (BDeu), and factorized Normalized Maximum Likelihood (fNML).

Main Results:

• Minimum Description Length (MDL/BIC) consistently outperformed other scoring functions in recovering true Bayesian network structures.
• Findings suggest MDL/BIC's superiority is due to the use of real-world data and optimal learning algorithms.
• Confirmed that larger sample sizes improve structure recovery accuracy and BDeu score sensitivity to parameter settings.

Conclusions:

• MDL/BIC is a highly effective scoring function for Bayesian network structure learning, especially with real-world data.
• The study validates previous findings on sample size effects and BDeu score sensitivity.
• Optimal learning algorithms and real-world data are crucial for reliable evaluation of scoring functions.